LLM-Enhanced Heterogeneous Graph Learning with Business Semantics for Financial Fraud Detection

Yixuan Chen; Hongyu Zhan; Jie Sheng; Weiyu Han; Shuai Chen; Tianyi Zhang; Xiao Tan; Jun Xia

cs.AIOct 26, 2026

LLM-Enhanced Heterogeneous Graph Learning with Business Semantics for Financial Fraud Detection

Authors: Yixuan Chen, Hongyu Zhan, Jie Sheng, Weiyu Han, Shuai Chen, Tianyi Zhang, Xiao Tan, Jun Xia

Organizations: HKUST-GZ & Jilin University · The Hong Kong University of Science and Technology (Guangzhou) · Ant International · HKUST-GZ & HKUST

Abstract

Financial fraud detection increasingly relies on heterogeneous transaction graphs, in which entities such as cards and merchants play distinct business roles and the risk meaning of an attribute depends on the entity's role: a high transaction volume may be routine for a merchant but unusual for a card. Heterogeneous graph neural networks preserve typed entities and relations but typically treat these attributes as role-agnostic inputs and leave their business meaning implicit. Meanwhile, most existing methods integrating large language models (LLMs) with graphs focus on homogeneous graphs or heterogeneous graphs with naturally available textual attributes, leaving their use on transaction graphs with structured attributes underexplored. To address this issue, we propose SemHGT, an LLM-enhanced framework that interprets structured attributes according to entity roles and makes the business semantics explicit in three stages: (1) role-aware prompts elicit risk-related business descriptions from node attributes, and a text encoder maps them to semantic vectors that enrich node representations; (2) semantic similarity guides multi-round sampling of target-specific path instances under transaction-to-transaction schemas; (3) schema-level tokens that combine local path evidence with global neighbor summaries are processed together with the target transaction by a unified transformer. Across three benchmark datasets, SemHGT leads the evaluated baselines on most dataset--metric comparisons. Further analyses provide additional evidence for the effectiveness of SemHGT's design and clarify how its semantic modeling and evidence selection contribute to fraud detection. Code is available at https://anonymous.4open.science/r/SemHGT-528A/.

Figures & tables

Figure 1. Overview of SemHGT. Role-aware business semantics enhance node representations and guide target-conditioned path sampling. For each schema, SemHGT fuses the sampled local path summary with the average representation of the target’s transaction neighbors under that schema to construct a schema-level token. A unified transformer jointly processes the target token and all schema-level tokens for fraud prediction. The SemHGT framework uses a language model and a frozen text encoder to derive role-aware business semantics from node attributes. These semantics enhance node features and guide path sampling. For each schema, the model fuses sampled local path evidence with the average representation of the target's transaction neighbors, then processes the resulting schema-level tokens with a target token using a unified transformer.

Figure 2. Global and local context for a target transaction. Global context averages the transaction neighbors under each schema, while local context captures the selected target-relevant path evidence. A target transaction is shown with two complementary contexts: its transaction neighbors under each schema and selected local path evidence around the target.

Dataset	Nodes	Edges	Metapaths
CreditCard	# node types: 6 # nodes: 1,860,300 Transaction: 1,852,394 Fraud: 9,651 (0.52%) Customer: 2,817 Card: 2,817 Merchant: 2,079 Category: 42 Location: 151	# edge types: 14 # edges: 18,535,208	$T$ –Customer– $T$ $T$ –Card– $T$ $T$ –Merchant– $T$ $T$ –Category– $T$ $T$ –Location– $T$
BankTrans	# node types: 6 # nodes: 8,987,098 Transaction: 8,914,963 Fraud: 13,332 (0.15%) Customer: 1,219 Card: 4,070 Merchant: 66,537 Category: 109 Location: 200	# edge types: 14 # edges: 89,227,542	$T$ –Customer– $T$ $T$ –Card– $T$ $T$ –Merchant– $T$ $T$ –Category– $T$ $T$ –Location– $T$
Elliptic++	# node types: 2 # nodes: 110,857 Transaction: 46,564 Fraud: 4,545 (9.76%) Wallet: 64,293	# edge types: 8 # edges: 402,882	$T$ – $T$ $T$ – $W$ – $T$ $T$ – $W$ – $W$ – $T$ $T$ – $W$ – $W$ – $W$ – $T$

Table 1. Statistics of the experimental graph views.

Dataset	Max epochs (patience)	Batch size	Learning rate
CreditCard	40 (10)	256	$10^{-3}$ (constant)
BankTrans	50 (10)	128	$10^{-4}\!\to\!10^{-5}$ (cosine)
Elliptic++	40 (10)	128	$3\times 10^{-4}$ (constant)

Table 2. Core dataset-specific training settings; early-stopping patience is shown in parentheses.

Method	CreditCard				BankTrans				Elliptic++
Method	Macro-F1	Fraud-F1	AUPRC	AUROC	Macro-F1	Fraud-F1	AUPRC	AUROC	Macro-F1	Fraud-F1	AUPRC	AUROC
MLP	72.41 $\mathbin{\pm}$ 0.15	44.99 $\mathbin{\pm}$ 0.31	40.83 $\mathbin{\pm}$ 0.24	94.29 $\mathbin{\pm}$ 0.14	54.47 $\mathbin{\pm}$ 0.67	9.17 $\mathbin{\pm}$ 1.41	3.76 $\mathbin{\pm}$ 0.02	92.75 $\mathbin{\pm}$ 0.02	79.10 $\mathbin{\pm}$ 0.61	60.31 $\mathbin{\pm}$ 1.20	59.57 $\mathbin{\pm}$ 0.59	84.49 $\mathbin{\pm}$ 0.41
SlotGAT ( Zhou et al., 2023 )	76.62 $\mathbin{\pm}$ 0.47	53.43 $\mathbin{\pm}$ 0.93	47.68 $\mathbin{\pm}$ 1.01	94.46 $\mathbin{\pm}$ 0.51	86.64 $\mathbin{\pm}$ 0.61	73.32 $\mathbin{\pm}$ 1.21	67.95 $\mathbin{\pm}$ 1.83	96.82 $\mathbin{\pm}$ 0.11	74.25 $\mathbin{\pm}$ 3.15	51.08 $\mathbin{\pm}$ 6.06	49.82 $\mathbin{\pm}$ 7.35	86.20 $\mathbin{\pm}$ 2.23
SimpleHGN ( Lv et al., 2021 )	63.31 $\mathbin{\pm}$ 0.89	26.89 $\mathbin{\pm}$ 1.78	14.75 $\mathbin{\pm}$ 0.40	84.80 $\mathbin{\pm}$ 0.14	83.93 $\mathbin{\pm}$ 0.65	67.92 $\mathbin{\pm}$ 1.29	55.35 $\mathbin{\pm}$ 6.37	96.02 $\mathbin{\pm}$ 0.20	74.29 $\mathbin{\pm}$ 0.49	51.55 $\mathbin{\pm}$ 0.91	52.30 $\mathbin{\pm}$ 0.74	87.56 $\mathbin{\pm}$ 0.35
FastGTN ( Yun et al., 2022 )	73.91 $\mathbin{\pm}$ 0.34	48.02 $\mathbin{\pm}$ 0.68	47.72 $\mathbin{\pm}$ 0.70	93.76 $\mathbin{\pm}$ 0.12	57.31 $\mathbin{\pm}$ 1.59	14.84 $\mathbin{\pm}$ 3.06	6.45 $\mathbin{\pm}$ 1.59	94.76 $\mathbin{\pm}$ 0.40	77.23 $\mathbin{\pm}$ 0.41	56.56 $\mathbin{\pm}$ 0.80	60.20 $\mathbin{\pm}$ 0.77	84.99 $\mathbin{\pm}$ 0.34
HAN ( Wang et al., 2019 )	83.34 $\mathbin{\pm}$ 0.88	66.81 $\mathbin{\pm}$ 1.75	70.29 $\mathbin{\pm}$ 1.72	96.97 $\mathbin{\pm}$ 0.34	80.25 $\mathbin{\pm}$ 0.42	60.55 $\mathbin{\pm}$ 0.84	54.74 $\mathbin{\pm}$ 2.47	94.35 $\mathbin{\pm}$ 0.41	71.71 $\mathbin{\pm}$ 0.22	47.58 $\mathbin{\pm}$ 0.42	55.34 $\mathbin{\pm}$ 1.37	85.30 $\mathbin{\pm}$ 0.21

Table 3. Main fraud-detection results on CreditCard, BankTrans, and Elliptic++. For each dataset, the table reports Macro-F1, Fraud-F1, AUPRC, and AUROC. All values are percentages, shown as mean

\pm

standard deviation across three runs. The best result in each metric column is boldfaced and shaded, and the second-best is underlined.

Figure 3. Ablation results on the full CreditCard test split for Macro-F1 and Fraud-F1 across six configurations combining local path evidence (

X

), global neighbor summaries (

Y

), and semantic enhancements. Dashed lines mark the

X

-only baseline, and annotations report the absolute percentage-point changes from this baseline for each configuration. Grouped bars compare Macro-F1 and Fraud-F1 across X, Y, X plus Y, X plus Y plus node-semantic enhancement, X plus Y plus semantic-guided local path sampling, and the full setting. X denotes local path evidence and Y global neighbor summaries. The full setting performs best on both metrics, and dashed lines mark the X-only baseline.

Figure 4. Hyperparameter sensitivity on the fixed, class-balanced test subsets of CreditCard (top) and Elliptic++ (bottom). From left to right, the columns vary candidate-pool size, context-node cap, schema-quota multiplier, and semantic-score weight

\lambda

. Orange circles show Macro-F1 (left axis), and blue triangles show AUPRC (right axis). Vertical dashed lines mark the reference setting for each parameter, while horizontal dashed lines indicate the corresponding metric scores. A two-row, four-column grid shows Macro-F1 and AUPRC versus candidate pool size, context node cap, schema quota multiplier, and semantic score weight for CreditCard and Elliptic++. Candidate-pool Macro-F1 generally rises over the displayed range, context and quota responses are nonmonotonic, and intermediate semantic weights yield higher AUPRC than endpoints.

Figure 5. Sensitivity to within-type semantic mismatch on the balanced CreditCard and Elliptic++ test subsets. The mismatch strength

\delta

controls the interpolation between each node’s semantic vector and that of a randomly permuted node of the same type. Lines show Macro-F1 and Fraud-F1 as semantic alignment decreases. Two vertically arranged line charts show Macro-F1 and Fraud-F1 against semantic mismatch strength

\delta

for CreditCard (top) and Elliptic++ (bottom). Both metrics decline as mismatch strength increases.

Dataset	Metric	$\delta=0$	$\delta=0.25$	$\delta=0.50$	$\delta=0.75$	$\delta=1$
CreditCard	Macro-F1	88.10	87.83	87.58	87.24	87.02
CreditCard	Fraud-F1	86.73	86.39	86.09	85.65	85.38
Elliptic++	Macro-F1	78.80	78.24	77.28	74.54	70.17
Elliptic++	Fraud-F1	74.64	73.81	72.50	68.62	62.22

Table 4. Macro-F1 and Fraud-F1 scores (%) under within-type semantic mismatch on the balanced CreditCard and Elliptic++ test subsets.

\delta

interpolates each node’s semantic vector with a same-type permuted vector, from original alignment (

\delta=0

) to the permuted vector (

\delta=1

Figure 6. Fraud-label composition of sampled terminal transaction endpoints. Panels show benign and fraudulent targets; colors distinguish datasets, shades distinguish sampling strategies, and annotations report semantic-minus-uniform differences (percentage points). Two panels compare fraud-endpoint rates under semantic-guided and uniform sampling across five transaction schemas for benign and fraudulent targets. Semantic-guided sampling generally lowers fraud-endpoint rates for benign targets and raises them for fraudulent targets.

Appendix figures & tables5 assets

Supplementary material from the paper’s appendix.

Appendix

Dataset	Split	Transactions	Fraud	Non-fraud
CreditCard	Train	1,111,436	6,401	1,105,035
	Validation	185,240	1,105	184,135
	Test	555,718	2,145	553,573
BankTrans	Train	891,496	1,333	890,163
	Validation	891,496	1,333	890,163
	Test	7,131,971	10,666	7,121,305

Appendix

Table 5. Transaction-label counts for each data split.

Dataset	Full validation	Subset (fraud / non-fraud)	Neg. weight
CreditCard	185,240	2,210 (1,105 / 1,105)	166.6380
BankTrans	891,496	22,661 (1,333 / 21,328)	41.7368
Elliptic++	4,005	550 (275 / 275)	13.5636

Appendix

Table 6. Validation subsets used for checkpoint selection and threshold calibration.

Dataset	Non-fraud	Fraud	Total
CreditCard	2,145	2,145	4,290
BankTrans	1,000	1,000	2,000
Elliptic++	868	868	1,736

Appendix

Table 7. Class-balanced test subsets used in analytical experiments other than ablation.

Dataset	Schema	Per-round path quota $k_{r}$	Max. context nodes
CreditCard	$T$ –Customer– $T$	8	112
	$T$ –Card– $T$	8
	$T$ –Merchant– $T$	8
	$T$ –Category– $T$	1
	$T$ –Location– $T$	1
BankTrans	$T$ –Customer– $T$	8	128

Appendix

Table 8. Local path schemas and sampling budgets.

T

denotes a transaction;

W

denotes a wallet.

Target label	Schema	Semantic	Uniform	$\Delta_{\mathrm{sem-unif}}$
CreditCard
Benign	same customer	0.179	0.371	-0.192
Benign	same card	0.173	0.371	-0.198
Benign	same merchant	0.207	0.386	-0.179
Benign	same category	0.245	0.379	-0.135
Benign	same location	0.193	0.423	-0.230

Appendix

Table 9. Fraud-endpoint rates for semantic-guided and uniform sampling. Rates are percentages;

\Delta

is semantic minus uniform in percentage points.

Financial Fraud Detection

Financial fraud detection identifies suspicious transactions, accounts, claims, and behavior to prevent losses and guide investigation.

See all 137 papers

May 27, 2026

Let Relations Speak: An End-to-End LLM-GNN Soft Prompt Framework for Fraud Detection

In recent years, Large Language Models (LLMs) have shown great capability in processing graph tasks such as fraud detection. However, most existing methods rely heavily on rich text attributes, which poses difficulties for this domain due to the lack of textual data. Although some pioneering methods attempt to overcome it, their textualization of graph structures via hard prompts easily leads to feature distortion. Additionally, fraud detection often exhibits multi-relational complexity, where current methods struggle to capture this deep semantic information. To address these challenges, we propose LLM-GNN Soft Prompt Framework (LGSPF). Specifically, LGSPF bridges the graph structure and semantic space using soft prompt to eliminate reliance on text. We further introduce a parallel Graph Neural Network (GNN) encoder to translate multi-relational topologies into graph tokens for fine-grained LLM fraud comprehension. Through end-to-end optimization, LGSPF enhances deep semantic alignment between LLM and GNN. Experiments across diverse fraud detection benchmarks demonstrate our method achieves state-of-the-art performance. Moreover, we further validate the contribution of LGSPF on enhancing the semantic interpretability of fraud behaviors.

Zhixing Zuo, Huilin He, Jiasheng Wu +1

School of Computer Science and Technology, Tongji University

Graph Neural Networks Graph Representation Learning Financial Fraud Detection Prompt Tuning LLMs for Graph Learning

Ask PDF

Oct 19, 2026

RGH-GNN : Risk-Gated Heterogeneous Graph Neural Network for Financial Fraud Detection

Credit card fraud is a significant problem for financial institutions, resulting in substantial monetary loss each year. Fraud patterns and methods are continuously evolving, with fraudulent transactions often disguised to resemble legitimate activity. Real world transaction data for fraud detection is severely affected by class imbalance making it harder to detect these transactions. Fraudulent transactions are rarely isolated. Accounts, devices, and other common attributes are often shared across fraudulent activity, revealing relational patterns that individual transaction features alone cannot capture. Graph-based approaches model these relationships explicitly. However, standard graph neural networks aggregate neighbourhood information uniformly, which can dilute the already scarce signal carried by fraudulent transactions and limit detection of camouflaged fraud. This work presents RGH-GNN, a heterogeneous graph neural network that addresses this limitation by incorporating transaction-level risk scores, generated by a supervised LightGBM classifier as edge weights. It is used to guide message passing through a learnable risk-gated aggregation mechanism. It ensures that high-risk transactions contribute more strongly to learned representations rather than being diluted by uniform neighbourhood aggregation. RGH-GNN is evaluated on the IEEE-CIS fraud detection dataset against standard graph neural network models that are trained on the same heterogeneous graph. RGH-GNN outperforms all baselines across every evaluation metric, with an AUC-ROC score of 0.914. This demonstrates that incorporating engineered risk signals into graph-based message passing substantially improves fraud detection performance.

Mythreyee Hari, Subin Sahayam, Vegesna S M Srinivasavarma

Department of Computer Science and Engineering, Shiv Nadar University, Chennai - 603110, Tamil Nadu, India

Financial Fraud Detection Graph Neural Networks Heterogeneous GNNs

Ask PDF

Sep 7, 2026

Temporal Heterogeneous Graph Transformer for Credit Card Fraud Detection

Credit card fraud detection typically relies on tabular features, while repeated attributes can also provide useful relational signals. This paper proposes THGT-FD, a Temporal Heterogeneous Graph Transformer for Fraud Detection. Each transaction is represented using one transaction token and six types of relation tokens and incorporates Time2Vec encoding into the transaction representation. A Transformer learns the interactions among these tokens within each individual transaction and then outputs a fraud probability. Experiments were conducted on 150,000 transactions sampled from the IEEE-CIS Fraud Detection dataset and chronologically partitioned according to TransactionDT. On the test set, THGT-FD achieved an AUC-ROC of 0.8536, an average precision of 0.4164, and a Recall@5% of 0.4708. The class-weighted histogram-based gradient-boosting baseline achieved an AUC-ROC of 0.8722. The results indicate that relation tokens provide useful information for fraud-risk ranking, although the current model does not yet incorporate entity-level historical aggregation.

Qinwen Yan

University of California, Los Angeles Los Angeles, California, USA

Financial Fraud Detection Transformer Feature Interaction Modeling

Ask PDF

Abstract

Financial Fraud Detection

Explore similar work

Let Relations Speak: An End-to-End LLM-GNN Soft Prompt Framework for Fraud Detection

RGH-GNN : Risk-Gated Heterogeneous Graph Neural Network for Financial Fraud Detection

Temporal Heterogeneous Graph Transformer for Credit Card Fraud Detection